Research Of Flame Object Detection Based On Convolutional Neural Network

Fire will cause great damage to the ecological environment,and seriously threaten the safety of human life and property,so how to detect fire timely and accurately is of great significance.Traditional fire sensors detect smoke,light,sound and infrared in a fire,but they are usually only used indoors and only detected when the fire is large.With the development of computer vision,image flame detection gradually replaces sensors and becomes the mainstream method of fire detection.According to different feature extraction methods,flame detection can be divided into artificial feature extraction and CNN(convolutional neural network)feature extraction.The latter has higher accuracy,faster speed and better scene generalization performance.According to different task types,flame detection can be divided into image classification and target detection.The former can not locate the flame position,can not reflect the fire scale,and is not sensitive to small-scale flame detection.Therefore,the flame object detection based on CNN is studied,the specific work is as follows:(1)Aiming at the problem that the anchor-based detection algorithm relies heavily on manual hyperparameter adjustment and is not suitable for objects with wide aspect ratio such as flame,the anchor-free algorithm FCOS is introduced and makes the following three improvements: firstly,the positive sample sampling method of center rectangle sampling is proposed to enhance the learning of object classification features;secondly,the Center-ness branch is replaced by the Io U(Intersection over Union)prediction branch on the detection subnetwork to improve the confidence score of the high-quality prediction box;finally,Soft-NMS is used instead of NMS(Non-Maximum Suppression)algorithm to prevent high-quality prediction box from being mistakenly screened.Experiments on the self-built flame image dataset show that the AP(Average Precision)of the improved FCOS is 2.6% higher than the original algorithm,and 0.4%higher than other detection algorithms.(2)Aiming at the difficulty of small object detection in early stage of fire,two small object augmentation methods,random copy-pasting small objects and Mosaic,are introduced to increase the number of small objects and enrich their context information in training,so that the network can focus more on the learning of small object features.Experimental results show that after using these two data augmentation methods in the above improved FCOS,AP and AP only on small objects(APs)are improved by 1.4%and 4.9% respectively,and are 2.1% and 1.9% higher than other detection algorithms..